Le Barrage de François Zola, Paul Cézanne, Public domain, via Wikimedia Commons
We are at an inflection point. Our shock and astonishment when ChatGPT was released to the public two years ago has given way to a realization that it will only be a short time until artificial intelligence saturates our lives. Some of us are impatient, anticipating medical breakthroughs and increased productivity. Others are fearful—for our jobs, most urgently—but also for the loss of human qualities that, until now, seemed untouchable.
In Ernest Hemingway’s “A Moveable Feast,” a memoir about his time as a young writer living in Paris, he describes a daily habit of taking himself to the museum after working on his book:
“In the afternoon I could walk through the gardens and then go to the Musee du Luxembourg where the great paintings were that have now mostly been transferred to the Louvre and the Jeu de Paume. I went there nearly every day for the Cézannes and to see the Manets and the Monets and the other impressionists that I had first come to know about in the Art Institute at Chicago. I was learning something from the painting of Cézanne that made writing simple true sentences far from enough to make the stories have the dimensions that I was trying to put in them. I was learning very much from him but I was not articulate enough to explain it to anyone. Besides, it was a secret.”
Ernest Hemingway 1923 passport photo, United States Government Printing Office., Public domain, via Wikimedia Commons
This is a remarkable passage—a window into the evolution of a genius—but it is also a warning about the risks of optimizing creativity.
Here, a young Hemingway is exploring his craft, tentatively feeling his way towards a new style. Day after day, he gazes at Cézanne’s brushstrokes—the image of a mountain emerging from tile-like planes of color, an intricate whole composed of simple, and similar, parts. Day after day, he notices the same subjects of rock and trees. They are in many of the paintings, but always fresh. Like his sentences, this is art that seems to him true and honest and real.
There is nothing surprising about someone in the 1920s confronting the novelty of Cézanne. The painter was modern then, even radical. What’s unexpected is that Hemingway turned to Cézanne as inspiration for his writing, discovering a path from color and pattern to narrative structure and the rhythm of words.
It worked. Hemingway’s pared-down style profoundly shaped 20th-century prose and influenced every writer who followed him. But the way he got there wasn’t obvious or efficient.
LLMs—Large Language Models—the AI engines behind chatbots like Google Gemini and Open AI’s ChatGPT, operate by figuring out the most statistically plausible next word in their training data, and then the next most plausible word, and then the next. For many purposes, this AI-generated content is sufficient—it’s conventional, it fits the task, and it is blazingly fast to produce. But an LLM’s objective in returning the most optimal result means that what we get is typical and often derivative.
La Montagne Sainte-Victoire vue de la carrière Bibémus, Paul Cézanne, Public domain, via Wikimedia Commons
That Hemingway refined his writing style by looking at a Cézanne painting was anything but typical. It would have been more efficient for him to find an example in another writer. But Hemingway was doing something new, not looking backwards, and not optimizing his creativity. He went to the Musee du Luxembourg and stood in front of a painting until a new way to write came together in his mind—so vague, at first, that he couldn’t even say what it was.
Throughout history, the most original thinkers and creators have often found inspiration outside their primary disciplines. Perhaps multimodal AI, machine learning models that work with many kinds of inputs, will mirror this trait. Results from these models are, in fact, beginning to show up, especially in science research. Or maybe experiments with chain-of-thought (CoT) prompting will introduce human-like reasoning. OpenAI-o1, launched on September 12 of this year is the first of a new CoT model series that is designed to “...spend more time thinking before they respond.”
In some circles, there is anticipation that AGI (artificial general intelligence) or ASI (artificial super intelligence) will integrate knowledge from any number of disciplines—perhaps *every* discipline. But even that may not be the key to elite human-level creativity.
Ernest Hemingway, Paris, 1924, John F. Kennedy Presidential Library and Museum, Boston, Public domain, via Wikimedia Commons
Hemingway was curious—and patient. He understood that new ideas emerge from unexpected places and take time to come together. He accepted that human thinkers get distracted, make wrong turns and are troubled by doubt. He worked hard at his craft but welcomed the daydream and a wander in the gardens.
Turn-by-turn GPS on our phones may have already claimed the ancient skill of navigation. It’s been at least two generations since electronic calculators reduced our desire and ability to play with numbers. And without regular practice, human traits like imagination, observation and curiosity will, likewise, get rusty and fall away.
As we increasingly turn to generative AI to help us create—the trait that separates our species from all others on this planet—we have to ask if an “optimal” process is really what we want. Especially from a tool that’s supposed to help.
